The Feasibility of Dynamically Granted Permissions: Aligning Mobile Privacy with User Preferences

Primal Wijesekera, Arjun Baokar, Lynn Tsai, Joel Reardon, Serge Egelman, David A. Wagner, Konstantin Beznosov

TL;DR

Mobile operating systems currently rely on "Ask-on-First-Use" (AOFU) permissions, which fail to account for how the context of a data request changes over time. This paper demonstrates a machine-learning approach that incorporates runtime context and passive behavioral traits to predict user privacy decisions with 96.8% accuracy, representing a four-fold reduction in error compared to current industry standards.

What Problem Does This Paper Address?

The "permission fatigue" vs. "privacy violation" trade-off is a central challenge in mobile OS design. Early Android used "Ask-on-Install" (AOI), an all-or-nothing approach that users rarely understood. Modern systems (Android M+, iOS) use "Ask-on-First-Use" (AOFU), where a user is prompted the first time an app requests a sensitive resource (like location).

The authors argue that AOFU is fundamentally flawed because it ignores contextual integrity. A user might be happy to share their location with a maps app while it's in the foreground, but they might object to that same app harvesting location data in the background for advertising. Current systems can't distinguish between these scenarios. If the system prompts too often, users become habituated and click "Allow" blindly; if it prompts too little, privacy is eroded.

Methods

The researchers took a multi-step approach to build a "Privacy Assistant":

  1. Instrumentation: They modified Android 5.1.1 to log 176 million events from 131 participants over a month. They tracked behavioral data (e.g., screen locking habits, HTTPS usage) and runtime context (e.g., which app is in the foreground).
  2. Ground Truth Collection: Using the Experience Sampling Method (ESM), the system occasionally interrupted users to ask if they would have allowed a specific, recent data request. This provided the "labels" for their ML model.
  3. Feature Engineering: They started with 16,000 candidate features and narrowed them down to 20 key predictors using Gini importance and Singular Value Decomposition (SVD).
  4. The Model: They employed a Support Vector Machine (SVM) with an RBF kernel.
  5. Learning Strategy: To solve the "cold start" problem for new users, they proposed a hybrid bootstrapping method. The system starts with a general model based on other users and then "learns" the specific user's nuances through a limited number of AOFU-style prompts.

Datasets

The study produced a unique longitudinal dataset:

  • Participants: 131 Android users (a mix of students, staff, and faculty).
  • Duration: Average of 32 days per participant.
  • Scale: 96 million sensitive permission requests and 4,224 verified user responses (ESM prompts).
  • Scope: 10 "dangerous" permission types, including location, contacts, and SMS.

Key Insights and Ideas

  • The "Contextual" vs. "Defaulter" Split: Users fall into two main camps. "Defaulters" (53%) are consistent—they almost always allow or always deny. "Contextuals" (47%) make nuanced decisions based on what they are doing at the moment.
  • Visibility is King: For "Contextuals," the visibility of the requesting app (foreground vs. background) was the strongest predictor of whether they would grant a permission.
  • Passive Traits Predict Privacy: Surprisingly, habits like how often a user manually locks their screen or their frequency of visiting HTTPS sites correlate with their privacy preferences. This allows the model to make educated guesses without even prompting the user.
  • AOFU is an "Upper Bound": The authors note that current AOFU implementations in the wild are likely less accurate than their study suggests because real systems often prompt for "check-only" API calls that don't actually access sensitive data, leading to even more false positives.

Key Results

  • High Accuracy: The final ML model achieved 96.8% accuracy in predicting user decisions.
  • Error Reduction: This is a 4x reduction in error compared to the standard AOFU model (84% accuracy) and a massive improvement over the old AOI model (25% accuracy).
  • Reduced User Burden: The system reached its "steady state" of high accuracy after only about 12 prompts per user.
  • Unbiased Predictions: The model's errors were split almost evenly between "privacy violations" (allowing when the user wanted to deny) and "functionality loss" (denying when the user wanted to allow), suggesting a balanced classifier.

Limitations and Future Work

  • The "No Consequences" Caveat: In this study, denying a permission didn't actually break the app (it was a simulation). In the real world, a user might "Allow" a request they dislike just to make the app work.
  • Adversarial ML: The authors acknowledge that if an OS used these features, malicious apps might try to "game" the system (e.g., by waiting for a specific foreground app to be active before requesting data).
  • Spurious Data: To prevent apps from crashing when the ML model denies a request, the authors suggest the OS should provide "spurious or artificial data" (like a mock location) instead of a hard denial.
  • UI/UX: Future work is needed to design a "Privacy Panel" where users can audit and override the decisions the ML model makes on their behalf.

Summary updated

Abstract

Current smartphone operating systems regulate application permissions by prompting users on an ask-on-first-use basis. Prior research has shown that this method is ineffective because it fails to account for context: the circumstances under which an application first requests access to data may be vastly different than the circumstances under which it subsequently requests access. We performed a longitudinal 131-person field study to analyze the contextuality behind user privacy decisions to regulate access to sensitive resources. We built a classifier to make privacy decisions on the user's behalf by detecting when context has changed and, when necessary, inferring privacy preferences based on the user's past decisions and behavior. Our goal is to automatically grant appropriate resource requests without further user intervention, deny inappropriate requests, and only prompt the user when the system is uncertain of the user's preferences. We show that our approach can accurately predict users' privacy decisions 96.8% of the time, which is a four-fold reduction in error rate compared to current systems.
BibTeX
@inproceedings{wijesekera2017feasibility,
  author = {Primal Wijesekera and Arjun Baokar and Lynn Tsai and Joel Reardon and Serge Egelman and David A. Wagner and Konstantin Beznosov},
  title = {{The Feasibility of Dynamically Granted Permissions: Aligning Mobile Privacy with User Preferences}},
  booktitle = {IEEE Symposium on Security and Privacy},
  year = {2017},
  url = {https://openreview.net/forum?id=aptRIvYMt6},
}

Loading related content...

Figure